The most powerful AI system Nvidia has ever built weighs as much as a pickup truck

Nvidia's Vera Rubin NVL72 redefines the unit of AI compute by packing 72 GPUs and 36 CPUs into a single liquid-cooled rack treated as one machine. The key enabler is NVLink 6, a proprietary interconnect running at 3.6 TB/s per GPU with direct chip-to-chip connectivity across all 72 GPUs simultaneously, eliminating the bottleneck of traditional PCIe buses. Liquid cooling makes the thermal density physically feasible. The result is 3.6 exaflops of compute and claimed 10x inference performance over Blackwell, signaling a shift where the rack — not the chip or server — becomes the fundamental unit of AI compute.